[ovirt-devel] Discussion VM CPU pinning and NUMA CPU pinning
Liang, Shang-Chun (David Liang, HPservers-Core-OE-PSC)
shangchun.liang at hp.com
Fri May 16 02:24:10 UTC 2014
Hi Gilad,
Thanks for your reply. Did you submit the UI code. We need to know your UI model to consolidate your feedback. Meanwhile, we need to know the current cpu pining solution that include the UI code you are in charging. Then, we will see if we can specific to RESTful.
Best Regards,
David Liang
> -----Original Message-----
> From: Gilad Chaplik [mailto:gchaplik at redhat.com]
> Sent: Thursday, May 15, 2014 8:07 PM
> To: Liao, Chuan (Jason Liao, HPservers-Core-OE-PSC)
> Cc: devel at ovirt.org; Doron Fediuck; Martin Sivák; Juan Hernandez; Malini
> Rao; Vinod, Chegu; Liang, Shang-Chun (David Liang, HPservers-Core-OE-PSC);
> Shi, Xiao-Lei (Bruce, HP Servers-PSC-CQ)
> Subject: Re: Discussion VM CPU pinning and NUMA CPU pinning
>
> Can you please be more specific on how to model RESTful API, with regards
> to CPU and CPU-pinning.
>
> Thanks,
> Gilad.
>
>
> ----- Original Message -----
> > From: "Chuan Liao (Jason Liao, HPservers-Core-OE-PSC)"
> <chuan.liao at hp.com>
> > To: devel at ovirt.org
> > Cc: "Doron Fediuck" <dfediuck at redhat.com>, "Gilad Chaplik"
> <gchaplik at redhat.com>, "Martin Sivák" <msivak at redhat.com>,
> > "Juan Hernandez" <juan.hernandez at redhat.com>, "Malini Rao"
> <mrao at redhat.com>, "Chegu Vinod" <chegu_vinod at hp.com>,
> > "Shang-Chun Liang (David Liang, HPservers-Core-OE-PSC)"
> <shangchun.liang at hp.com>, "Xiao-Lei Shi (Bruce, HP
> > Servers-PSC-CQ)" <xiao-lei.shi at hp.com>
> > Sent: Thursday, May 15, 2014 2:50:23 PM
> > Subject: Discussion VM CPU pinning and NUMA CPU pinning
> >
> > Hi All,
> >
> > Now we are working on the NUMA tune feature development for oVirt 3.5.
> This
> > feature allow user to configure the vCPU pining according to host
> CPU/NUMA
> > topology to get the best performance for created VM. But this will impact
> > the current function of vCPU pining. Now, we are looking for a solution to
> > consolidate the conflict. We want to start an open discussion about VM
> CPU
> > pinning (the current design) and NUMA CPU pining (new feature we are
> working
> > on) in developer list:
> >
> > Background:
> > Concept:
> >
> > 1. VM CPU pinning: the exist feature in current oVirt, which allow user
> > to configure the VM vCPU pinning over ovirt, user can configure vCPU
> pining
> > independently without NUMA tune.
> >
> > 2. NUMA CPU pinning: allow user to configure the vCPU pining according
> > to host NUMA topology to get the best performance for created VM
> >
> > 3. vNode: Virtual NUMA node ( User configured )
> >
> > 4. pNode: host physical NUMA node ( Get capability from host )
> > Notice:
> >
> > 1. NUMA tuning feature and CPU pinning feature are individually in
> > libvirt ( ovirt backend ).
> >
> > 2. User could configure VM CPU pining individually without NUMA tune
> > setup.
> >
> > 3. When configuring NUMA tuning feature, user need to configure VM
> NUMA
> > tuning ( vNode pinto pNode & tuning mode ) and VM CPU pinning ( NUMA
> > included ) to get optimized VM performance, otherwise VM will have low
> > performance.
> >
> > We have two proposal now for this issue. Please give us some comments
> and
> > your feedback, thanks.
> >
> > Solution 1:
> > GUI:
> > Transform between the CPU pinning text and a structure which will be
> used in
> > NUMA CPU pinning configure page.
> > And then save the CPU pinning text to the current VM CPU pinning field
> of
> > VM.
> > Restful:
> > Transform between the CPU pinning text and a structure which will be
> used in
> > restful NUMA CPU pinning.
> > And then save the CPU pinning text to the current VM CPU pinning field of
> VM.
> > Broker:
> > Remove temporary solution(See the current implementation as below)
> and follow
> > previous cpupin configuration procedure.
> >
> > Solution 2:
> > GUI:
> > If current VM CPU pinning is configured, when user open NUMA CPU
> pinning
> > configure page, he will get a warning message. If he continues to configure
> > NUMA CPU pinning and save the data, the current VM CPU pinning
> > configuration will be cleared.
> > NUMA CPU pinning configured data will be saved as new structure
> individually
> > without changing current VM CPU pinning configuration.
> > Restful:
> > Configure NUMA CPU pinning and then save the data with the new NUMA
> CPU
> > pinning structure.
> > Entity and Database:
> > Individually NUMA CPU pinning entities and data structure.
> > Broker:
> > NUMA CPU pinning configuration will be first considered to use. If
> > it's not configured, it will use the current VM CPU pinning.
> >
> > Solution 1 The CPU pining data is consistent but the code logic is very
> > complex.
> > Solution 2 have better adaptive and better data structure. This is the way
> we
> > preferred.
> >
> > We are appreciated that anybody could give us your comments or the
> better
> > solution you have.
> >
> >
> >
> > The below is the current implementation for your reference:
> > VM CPU pinning feature
> > GUI:
> > User input vCPU pining configure data with formatted text.
> > Restful:
> > User configure CpuTune and VCpuPin model with mapper to CPU pinning
> text.
> > Entity and Database:
> > CPU pinning is String property.
> > Broker:
> > Generate structure of cputune ( libvirt format ) from CPU pining string
> > property.
> > NUMA tuning feature
> > GUI:
> > User can drag & drop virtual NUMA node to host NUMA node ( pin to or
> remove
> > pin to ).
> > Restful:
> > Add/update/remove virtual NUMA node with property of pin to host
> NUMA node
> > index.
> > Design NUMACPU model under NUMA node for extend.
> > Entity and Database:
> > Individually NUMA node entities ( vNode extend pNode ) and store
> procedure.
> > Broker:
> > Generate structure of numatune, cpu/numa ( libvirt format ) from NUMA
> node
> > entities.
> > Temporary solution prevent Notice 3
> > Broker:
> > Generate right structure of cputune ( libivirt format ) from NUMA node
> > entities ( vNode pinto pNode )
> > Limitation:
> > This structure of cputune will not get the best performance of vNode.
> >
> > Best Regards,
> > Jason Liao
> >
> >
More information about the Devel
mailing list